QUESTION:
Please explain the notion of TRANSFER FUNCTION
identification
and the role of the pre-whitening filter as a tool in this
step ?
ANSWER:
Consider a Transfer Function of the form:
Y(t ) = V0 X(t ) + V1 X(t-1)
+ V2 X(t-2) .... + Vj X(t-j) + N(t )
or Y(t ) = [W(B)/D(B)] X(t ) +
[T(B)/P(B)] A(t )
where an ARIMA model for X exists such that
E(t ) = [T1(B)/P1(B)] X(t )
equivalently
X(t ) = [P1(B)/T1(B)] E(t )
Now more generally our Transfer Function may contain a pure
delay of b periods
Y(t ) = [W(B)/D(B)] X(t-b) +
[T(B)/P(B)] A(t )
where
b = the number of periods of pure delay before Y responds
to X
{T(B)/P(B)} = ARMA model for unobserved series A
W(b) = input lag structure reflecting static relationship
of
Y to X and is a polynomial of order s-1
D(b) = output lag structure reflecting dynamic relationship
of
Y to X and is a polynomial of order r
and substituting the stationary y and Y for x and X we get:
y(t) = [W(B)/D(B)] x(t-b) +
[T(B)/P(B)] A(t) [1]
If we now identify the following ARMA for x
E(t) = [T1(B)/P1(B)] x(t) [2]
If we now multiply the equation [1] above by [T1(B)/P1(B)]
[T1(B)/P1(B)]y(t ) = [W(B)/D(B)][T1(B)/P1(B)] x(t-b)
+ [T1(B)/P1(B)][T(B)/P(B)] A(t )
Substituting [2] into the current equation generates
[T1(B)/P1(B)]y(t ) = [W(B)/D(B)] E(t-b)
+ [T1(B)/P1(B)][T(B)/P(B)] A(t)
If we let
y(t )=[T1(B)/P1(B)]y(t ) and aa(t
)=[T1(B)/P1(B)][T(B)/P(B)] A(t )
we have
yy(t ) = [W(B)/D(B)] E(t-b) + aa(t )
where E(t ) is N.I.I.D. Thus we get a clear
picture
of [W(B)/D(B)] which represents the Transfer of X into Y.
In practice we identify [W(B)/D(B)] via the cross-correlations
of
yy(t ) and E(t ) , selecting orders
of [W(B)/D(B)] sufficient to
describe the relationship. The pure delay (b) represents
the observed
dead time. Upon identifying [W(B)/D(B)] we can then back
out the effect
of X on Y and obtain an initial estimate of N(t
) from:
Y(t ) = V0 X(t ) + V1 X(t-1)
+ V2 X(t-2) .... + Vj X(t-j) + N(t )
Y(t) - V0 X(t) - V1 X(t-1)
- V2 X(t-2) .... - Vj X(t-j) = N(t )
We can now study N(t) and obtain the two
polynomials T(B) and P(B).
N(t) = [T(B)/P(B)] A(t) which
completes the identification step.
This explains the need for filtering and exposes the
relationship
of the prewhitening filter in identifying the form of the
inter-relationship
between the observed series Y and X.
Y(t) = [W(B)/D(B)] X(t) + [T(B)/P(B)]
A(t)